XpertAI: uncovering regression model strategies for sub-manifolds

Regression models need different explanation approaches than classification models: The paper identifies that existing Explainable AI methods focus primarily on classification tasks, but regression models require more precise explanations tailored to specific questions (e.g., explaining why a prediction is above 50 rather than just above 0).

XpertAI breaks down model behavior into range-specific strategies: The framework divides how a regression model makes predictions into multiple sub-strategies that apply to different output ranges, allowing users to understand which factors drive predictions in specific ranges rather than treating the model as a single monolithic decision-maker.

The approach works with existing explanation techniques: Rather than requiring entirely new methods, XpertAI is designed to work alongside popular existing XAI approaches (occlusion, gradient integration, and reverse propagation), making it potentially easier to integrate into current AI systems.

Enables more targeted queries about model behavior: Users can now formulate precise questions about regression models as combinations of these sub-strategies, moving beyond generic explanations to understand the specific factors influencing predictions in particular output ranges.

Demonstrates practical improvements: The paper provides both qualitative and quantitative evidence that this approach offers benefits, suggesting the framework could help practitioners better validate and understand regression models used in real-world applications.